diff options
author | Linfeng Zhang <linfengz@google.com> | 2017-01-06 16:47:22 +0000 |
---|---|---|
committer | Gerrit Code Review <noreply-gerritcodereview@google.com> | 2017-01-06 16:47:23 +0000 |
commit | 2d12a52ff09ff47cd69659b6d2f862af1913c779 (patch) | |
tree | 5f27782a39fc5bb2e66f94bf66c36b928e5122fa /vpx_dsp/arm/idct_neon.h | |
parent | 90f889a56d14b9335d6710af9d4034c101927b3b (diff) | |
parent | 9b187954df6044fcc533b1ad43761f3f8cdf8b1a (diff) | |
download | libvpx-2d12a52ff09ff47cd69659b6d2f862af1913c779.tar libvpx-2d12a52ff09ff47cd69659b6d2f862af1913c779.tar.gz libvpx-2d12a52ff09ff47cd69659b6d2f862af1913c779.tar.bz2 libvpx-2d12a52ff09ff47cd69659b6d2f862af1913c779.zip |
Merge "Add high bitdepth 8x8 idct NEON intrinsics"
Diffstat (limited to 'vpx_dsp/arm/idct_neon.h')
-rw-r--r-- | vpx_dsp/arm/idct_neon.h | 245 |
1 files changed, 245 insertions, 0 deletions
diff --git a/vpx_dsp/arm/idct_neon.h b/vpx_dsp/arm/idct_neon.h index 04352c491..aecc543db 100644 --- a/vpx_dsp/arm/idct_neon.h +++ b/vpx_dsp/arm/idct_neon.h @@ -24,6 +24,13 @@ DECLARE_ALIGNED(16, static const int16_t, kCospi[8]) = { -9102 /* -cospi_20_64 */, 3196 /* cospi_28_64 */ }; +DECLARE_ALIGNED(16, static const int32_t, kCospi32[8]) = { + 16384 /* cospi_0_64 */, 15137 /* cospi_8_64 */, + 11585 /* cospi_16_64 */, 6270 /* cospi_24_64 */, + 16069 /* cospi_4_64 */, 13623 /* cospi_12_64 */, + -9102 /* -cospi_20_64 */, 3196 /* cospi_28_64 */ +}; + //------------------------------------------------------------------------------ // Helper functions used to load tran_low_t into int16, narrowing if necessary. @@ -217,4 +224,242 @@ static INLINE void idct4x4_16_kernel_bd8(const int16x4_t cospis, *a1 = vsubq_s16(d0, d1); } +static INLINE void idct8x8_12_pass1_bd8( + const int16x4_t cospis0, const int16x4_t cospisd0, const int16x4_t cospisd1, + int16x4_t *const io0, int16x4_t *const io1, int16x4_t *const io2, + int16x4_t *const io3, int16x4_t *const io4, int16x4_t *const io5, + int16x4_t *const io6, int16x4_t *const io7) { + int16x4_t step1[8], step2[8]; + int32x4_t t32[2]; + + transpose_s16_4x4d(io0, io1, io2, io3); + + // stage 1 + step1[4] = vqrdmulh_lane_s16(*io1, cospisd1, 3); + step1[5] = vqrdmulh_lane_s16(*io3, cospisd1, 2); + step1[6] = vqrdmulh_lane_s16(*io3, cospisd1, 1); + step1[7] = vqrdmulh_lane_s16(*io1, cospisd1, 0); + + // stage 2 + step2[1] = vqrdmulh_lane_s16(*io0, cospisd0, 2); + step2[2] = vqrdmulh_lane_s16(*io2, cospisd0, 3); + step2[3] = vqrdmulh_lane_s16(*io2, cospisd0, 1); + + step2[4] = vadd_s16(step1[4], step1[5]); + step2[5] = vsub_s16(step1[4], step1[5]); + step2[6] = vsub_s16(step1[7], step1[6]); + step2[7] = vadd_s16(step1[7], step1[6]); + + // stage 3 + step1[0] = vadd_s16(step2[1], step2[3]); + step1[1] = vadd_s16(step2[1], step2[2]); + step1[2] = vsub_s16(step2[1], step2[2]); + step1[3] = vsub_s16(step2[1], step2[3]); + + t32[1] = vmull_lane_s16(step2[6], cospis0, 2); + t32[0] = vmlsl_lane_s16(t32[1], step2[5], cospis0, 2); + t32[1] = vmlal_lane_s16(t32[1], step2[5], cospis0, 2); + step1[5] = vrshrn_n_s32(t32[0], 14); + step1[6] = vrshrn_n_s32(t32[1], 14); + + // stage 4 + *io0 = vadd_s16(step1[0], step2[7]); + *io1 = vadd_s16(step1[1], step1[6]); + *io2 = vadd_s16(step1[2], step1[5]); + *io3 = vadd_s16(step1[3], step2[4]); + *io4 = vsub_s16(step1[3], step2[4]); + *io5 = vsub_s16(step1[2], step1[5]); + *io6 = vsub_s16(step1[1], step1[6]); + *io7 = vsub_s16(step1[0], step2[7]); +} + +static INLINE void idct8x8_12_pass2_bd8( + const int16x4_t cospis0, const int16x4_t cospisd0, const int16x4_t cospisd1, + const int16x4_t input0, const int16x4_t input1, const int16x4_t input2, + const int16x4_t input3, const int16x4_t input4, const int16x4_t input5, + const int16x4_t input6, const int16x4_t input7, int16x8_t *const output0, + int16x8_t *const output1, int16x8_t *const output2, + int16x8_t *const output3, int16x8_t *const output4, + int16x8_t *const output5, int16x8_t *const output6, + int16x8_t *const output7) { + int16x8_t in[4]; + int16x8_t step1[8], step2[8]; + int32x4_t t32[8]; + int16x4_t t16[8]; + + transpose_s16_4x8(input0, input1, input2, input3, input4, input5, input6, + input7, &in[0], &in[1], &in[2], &in[3]); + + // stage 1 + step1[4] = vqrdmulhq_lane_s16(in[1], cospisd1, 3); + step1[5] = vqrdmulhq_lane_s16(in[3], cospisd1, 2); + step1[6] = vqrdmulhq_lane_s16(in[3], cospisd1, 1); + step1[7] = vqrdmulhq_lane_s16(in[1], cospisd1, 0); + + // stage 2 + step2[1] = vqrdmulhq_lane_s16(in[0], cospisd0, 2); + step2[2] = vqrdmulhq_lane_s16(in[2], cospisd0, 3); + step2[3] = vqrdmulhq_lane_s16(in[2], cospisd0, 1); + + step2[4] = vaddq_s16(step1[4], step1[5]); + step2[5] = vsubq_s16(step1[4], step1[5]); + step2[6] = vsubq_s16(step1[7], step1[6]); + step2[7] = vaddq_s16(step1[7], step1[6]); + + // stage 3 + step1[0] = vaddq_s16(step2[1], step2[3]); + step1[1] = vaddq_s16(step2[1], step2[2]); + step1[2] = vsubq_s16(step2[1], step2[2]); + step1[3] = vsubq_s16(step2[1], step2[3]); + + t32[2] = vmull_lane_s16(vget_low_s16(step2[6]), cospis0, 2); + t32[3] = vmull_lane_s16(vget_high_s16(step2[6]), cospis0, 2); + t32[0] = vmlsl_lane_s16(t32[2], vget_low_s16(step2[5]), cospis0, 2); + t32[1] = vmlsl_lane_s16(t32[3], vget_high_s16(step2[5]), cospis0, 2); + t32[2] = vmlal_lane_s16(t32[2], vget_low_s16(step2[5]), cospis0, 2); + t32[3] = vmlal_lane_s16(t32[3], vget_high_s16(step2[5]), cospis0, 2); + t16[0] = vrshrn_n_s32(t32[0], 14); + t16[1] = vrshrn_n_s32(t32[1], 14); + t16[2] = vrshrn_n_s32(t32[2], 14); + t16[3] = vrshrn_n_s32(t32[3], 14); + step1[5] = vcombine_s16(t16[0], t16[1]); + step1[6] = vcombine_s16(t16[2], t16[3]); + + // stage 4 + *output0 = vaddq_s16(step1[0], step2[7]); + *output1 = vaddq_s16(step1[1], step1[6]); + *output2 = vaddq_s16(step1[2], step1[5]); + *output3 = vaddq_s16(step1[3], step2[4]); + *output4 = vsubq_s16(step1[3], step2[4]); + *output5 = vsubq_s16(step1[2], step1[5]); + *output6 = vsubq_s16(step1[1], step1[6]); + *output7 = vsubq_s16(step1[0], step2[7]); +} + +static INLINE void idct8x8_64_1d_bd8(const int16x4_t cospis0, + const int16x4_t cospis1, + int16x8_t *const io0, int16x8_t *const io1, + int16x8_t *const io2, int16x8_t *const io3, + int16x8_t *const io4, int16x8_t *const io5, + int16x8_t *const io6, + int16x8_t *const io7) { + int16x4_t input_1l, input_1h, input_3l, input_3h, input_5l, input_5h, + input_7l, input_7h; + int16x4_t step1l[4], step1h[4]; + int16x8_t step1[8], step2[8]; + int32x4_t t32[8]; + int16x4_t t16[8]; + + transpose_s16_8x8(io0, io1, io2, io3, io4, io5, io6, io7); + + // stage 1 + input_1l = vget_low_s16(*io1); + input_1h = vget_high_s16(*io1); + input_3l = vget_low_s16(*io3); + input_3h = vget_high_s16(*io3); + input_5l = vget_low_s16(*io5); + input_5h = vget_high_s16(*io5); + input_7l = vget_low_s16(*io7); + input_7h = vget_high_s16(*io7); + step1l[0] = vget_low_s16(*io0); + step1h[0] = vget_high_s16(*io0); + step1l[1] = vget_low_s16(*io2); + step1h[1] = vget_high_s16(*io2); + step1l[2] = vget_low_s16(*io4); + step1h[2] = vget_high_s16(*io4); + step1l[3] = vget_low_s16(*io6); + step1h[3] = vget_high_s16(*io6); + + t32[0] = vmull_lane_s16(input_1l, cospis1, 3); + t32[1] = vmull_lane_s16(input_1h, cospis1, 3); + t32[2] = vmull_lane_s16(input_3l, cospis1, 2); + t32[3] = vmull_lane_s16(input_3h, cospis1, 2); + t32[4] = vmull_lane_s16(input_3l, cospis1, 1); + t32[5] = vmull_lane_s16(input_3h, cospis1, 1); + t32[6] = vmull_lane_s16(input_1l, cospis1, 0); + t32[7] = vmull_lane_s16(input_1h, cospis1, 0); + t32[0] = vmlsl_lane_s16(t32[0], input_7l, cospis1, 0); + t32[1] = vmlsl_lane_s16(t32[1], input_7h, cospis1, 0); + t32[2] = vmlal_lane_s16(t32[2], input_5l, cospis1, 1); + t32[3] = vmlal_lane_s16(t32[3], input_5h, cospis1, 1); + t32[4] = vmlsl_lane_s16(t32[4], input_5l, cospis1, 2); + t32[5] = vmlsl_lane_s16(t32[5], input_5h, cospis1, 2); + t32[6] = vmlal_lane_s16(t32[6], input_7l, cospis1, 3); + t32[7] = vmlal_lane_s16(t32[7], input_7h, cospis1, 3); + t16[0] = vrshrn_n_s32(t32[0], 14); + t16[1] = vrshrn_n_s32(t32[1], 14); + t16[2] = vrshrn_n_s32(t32[2], 14); + t16[3] = vrshrn_n_s32(t32[3], 14); + t16[4] = vrshrn_n_s32(t32[4], 14); + t16[5] = vrshrn_n_s32(t32[5], 14); + t16[6] = vrshrn_n_s32(t32[6], 14); + t16[7] = vrshrn_n_s32(t32[7], 14); + step1[4] = vcombine_s16(t16[0], t16[1]); + step1[5] = vcombine_s16(t16[2], t16[3]); + step1[6] = vcombine_s16(t16[4], t16[5]); + step1[7] = vcombine_s16(t16[6], t16[7]); + + // stage 2 + t32[2] = vmull_lane_s16(step1l[0], cospis0, 2); + t32[3] = vmull_lane_s16(step1h[0], cospis0, 2); + t32[4] = vmull_lane_s16(step1l[1], cospis0, 3); + t32[5] = vmull_lane_s16(step1h[1], cospis0, 3); + t32[6] = vmull_lane_s16(step1l[1], cospis0, 1); + t32[7] = vmull_lane_s16(step1h[1], cospis0, 1); + t32[0] = vmlal_lane_s16(t32[2], step1l[2], cospis0, 2); + t32[1] = vmlal_lane_s16(t32[3], step1h[2], cospis0, 2); + t32[2] = vmlsl_lane_s16(t32[2], step1l[2], cospis0, 2); + t32[3] = vmlsl_lane_s16(t32[3], step1h[2], cospis0, 2); + t32[4] = vmlsl_lane_s16(t32[4], step1l[3], cospis0, 1); + t32[5] = vmlsl_lane_s16(t32[5], step1h[3], cospis0, 1); + t32[6] = vmlal_lane_s16(t32[6], step1l[3], cospis0, 3); + t32[7] = vmlal_lane_s16(t32[7], step1h[3], cospis0, 3); + t16[0] = vrshrn_n_s32(t32[0], 14); + t16[1] = vrshrn_n_s32(t32[1], 14); + t16[2] = vrshrn_n_s32(t32[2], 14); + t16[3] = vrshrn_n_s32(t32[3], 14); + t16[4] = vrshrn_n_s32(t32[4], 14); + t16[5] = vrshrn_n_s32(t32[5], 14); + t16[6] = vrshrn_n_s32(t32[6], 14); + t16[7] = vrshrn_n_s32(t32[7], 14); + step2[0] = vcombine_s16(t16[0], t16[1]); + step2[1] = vcombine_s16(t16[2], t16[3]); + step2[2] = vcombine_s16(t16[4], t16[5]); + step2[3] = vcombine_s16(t16[6], t16[7]); + + step2[4] = vaddq_s16(step1[4], step1[5]); + step2[5] = vsubq_s16(step1[4], step1[5]); + step2[6] = vsubq_s16(step1[7], step1[6]); + step2[7] = vaddq_s16(step1[7], step1[6]); + + // stage 3 + step1[0] = vaddq_s16(step2[0], step2[3]); + step1[1] = vaddq_s16(step2[1], step2[2]); + step1[2] = vsubq_s16(step2[1], step2[2]); + step1[3] = vsubq_s16(step2[0], step2[3]); + + t32[2] = vmull_lane_s16(vget_low_s16(step2[6]), cospis0, 2); + t32[3] = vmull_lane_s16(vget_high_s16(step2[6]), cospis0, 2); + t32[0] = vmlsl_lane_s16(t32[2], vget_low_s16(step2[5]), cospis0, 2); + t32[1] = vmlsl_lane_s16(t32[3], vget_high_s16(step2[5]), cospis0, 2); + t32[2] = vmlal_lane_s16(t32[2], vget_low_s16(step2[5]), cospis0, 2); + t32[3] = vmlal_lane_s16(t32[3], vget_high_s16(step2[5]), cospis0, 2); + t16[0] = vrshrn_n_s32(t32[0], 14); + t16[1] = vrshrn_n_s32(t32[1], 14); + t16[2] = vrshrn_n_s32(t32[2], 14); + t16[3] = vrshrn_n_s32(t32[3], 14); + step1[5] = vcombine_s16(t16[0], t16[1]); + step1[6] = vcombine_s16(t16[2], t16[3]); + + // stage 4 + *io0 = vaddq_s16(step1[0], step2[7]); + *io1 = vaddq_s16(step1[1], step1[6]); + *io2 = vaddq_s16(step1[2], step1[5]); + *io3 = vaddq_s16(step1[3], step2[4]); + *io4 = vsubq_s16(step1[3], step2[4]); + *io5 = vsubq_s16(step1[2], step1[5]); + *io6 = vsubq_s16(step1[1], step1[6]); + *io7 = vsubq_s16(step1[0], step2[7]); +} + #endif // VPX_DSP_ARM_IDCT_NEON_H_ |